Skip to content
Faisal Mahmud TusherDigital marketing, SEO & analytics
with Faisal Tusher

AI in Marketing › Troubleshooting

Customer Data Pasted Into AI Tools

Why customer records keep ending up in AI tools: no data rule, no approved business-tier tool, unprepared exports, agencies' own tools and training on demos.
Troubleshootingintermediate12 min readUpdated

First, tell the data protection owner, then list every case by material, tool, person and task

Before any diagnosis, tell the person who owns data protection in the business, today. The AI governance operations playbook gives that person the decision on which personal data may reach each tool, and on anything that follows when an incident involves personal data, and the tutorial on writing an AI use policy counts personal data entered into an unapproved tool as an incident. Whether anything must be deleted, recorded or reported is their decision, and this page makes no judgement on it. Say plainly to the team that reporting a case is welcome and will not be confused with the case itself, because a case people hide cannot be listed.

Then list the cases, each with four entries: the material (which file, and which fields it carried), the tool and plan (free or paid, personal or business account, or a feature inside an existing platform), the person (staff, agency or freelancer) and the task. Ask people to add their own cases, and read the admin logs where an approved tool keeps them. Pass the list to the data protection owner as it grows.

Read the columns, not the names: free plans point at the missing approved tool, whole exports at the missing routine, unfamiliar tools at the missing approval route, agency names at the supplier gap and platform names at features switched on under changed terms. The NIST AI Risk Management Framework, as the framework page applies it, orders the fixes: govern (a rule, owners and approved tools), map (where AI is used and what data each use sees), measure (checks on what goes in) and manage (changing the routine where the checks find gaps).

This page covers the habit. Output that reached customers belongs to the guide to handling an AI mistake, an automated send that went wrong to the diagnosis of wrong messages sent by an AI automation, and figures nobody can reproduce to the diagnosis for AI report numbers that do not match.

Diagnosis table

Diagnosis

SymptomCheckLikely causeFix
People give different answers about what may go into an AI toolAsk a few people to state the rule unpromptedNo data rule short enough to rememberWrite the three-class rule in one paragraph beside the tools
Cases sit on free or personal accountsThe plan behind each case and its written termsNo approved business-tier tool for the taskApprove a tool on a plan whose written terms meet the data rule
Cases are whole exports, full transcripts or entire ticket threadsThe fields pasted against the fields the task neededWhole exports pasted because preparing them takes longerSet a minimisation routine with saved export views and a sample read
Tools in the list appear nowhere in the policyWhether the policy names tools, plans and an approverA policy that never named the tools, the uses and the approverBuild the register of uses, name tools and owners, publish the route for a new tool
Agency or freelancer names appear in the person columnWhether supplier agreements carry the data ruleAgencies and freelancers using tools nobody evaluatedExtend the policy and data rule to suppliers in writing
Customer records reach an AI feature inside an existing platformWhen the feature was switched on, by whom, under which termsAI features in existing platforms under terms that changed when the feature arrivedRead the feature's terms and switch it off until they meet the rule
People can recite the rule but paste exports on real tasksWhether training used the team's own material or a vendor demoTraining on demos rather than the team's own tasksPractise on real tasks, including preparing the material and choosing the tool

Fix, verify and prevent each cause

No data rule short enough to remember

The disclosure and data privacy guide calls this the commonest privacy failure in marketing AI, and a mundane one: someone pastes a list or transcript to improve the output, and the terms of the tool let the vendor retain and learn from it. Its remedy sorts material into three classes: public or non-personal material may enter any approved tool; personal data, meaning anything about an identifiable person, enters only a tool whose contract rules out training on inputs, caps retention and fixes a region, and only when the task requires it; special-category details such as health stay out of marketing prompts altogether.

Fix: write the three classes in one paragraph in the team's own words, beside the tools, with a named person to ask when material does not fit.

Verify: a week later, ask a few people to state the rule and sort a handful of real files into its classes.

Prevent: repeat the recall check at each policy review, and treat a rule people cannot state as too long.

No approved business-tier tool for the task

With no approved tool for a task, people use whatever is open, often a consumer plan. The disclosure guide notes that one product's consumer and business plans frequently come with different terms, and business tiers usually answer the data questions acceptably where consumer tiers do not. The guide to evaluating AI tools collects written answers, for the plan you would buy, on what the vendor does about training, retention, sub-processors, region and its own staff's access, and on who owns the output. Where a vendor processes personal data for the business it acts as a processor under the GDPR, and the guide describes what that contract covers; a vendor unable to offer those terms is ruled out for personal data.

Fix: for each task on the case list, approve a tool on a plan whose written terms meet the rule, or state that the task uses only non-personal material.

Verify: confirm training on inputs is off by contract rather than by a settings toggle, and that the admin console enforces it for every account.

Prevent: keep approved tools and their plans on one list beside the rule, and recheck it when a vendor changes its plans or terms.

Whole exports pasted because preparing them takes longer than not

The routine for minimising data before prompting starts from this habit: exports of First-Party Data tend to leave with every field attached, since trimming them costs extra clicks. Whatever reaches the tool has been processed, so the decision about what to share comes before the paste. The GDPR, as the tutorial reads it, lists data minimisation among its principles and expects processing for specified purposes, and the ICO's direct marketing guidance sets out how the UK's data protection and electronic marketing rules bear on marketing work.

Fix: make the routine quicker than the shortcut: a one-sentence task, only the fields it needs, dropped in the export settings rather than afterwards, aggregation where rows are not needed, pseudonymised identifiers with the key kept elsewhere, masked free text and a sample read. Saved export views remove most of the clicks.

Verify: compare the fields in the next prepared files with the task sentence, and read the messiest records for sign-offs and health details the patterns missed.

Prevent: record each preparation beside the use's register row, add missed identifiers to the pattern list, and remember that pseudonymised data is still personal data.

A policy that never named the tools, the uses and who approves a new one

The policy tutorial begins with a register of the uses already in place, built from each person's list and a check of the stack, because built-in features are routinely missed. From it an AI use policy lists each approved tool and the plan it is approved on, with an owner for every use, and publishes the route for approving a new tool, with nothing above the internal tier running on an unapproved one.

Fix: turn the case list into the first register rows, stop any use that needs stopping now and note it, and write the approval route in a sentence.

Verify: walk a request for one listed tool through the route to the named person who checks its terms.

Prevent: give the register an owner, add a row before a new use starts, and bring the policy review forward for a new tool, an incident or a use changing tier.

Agencies and freelancers using tools nobody evaluated

The guide to setting AI rules for agencies explains why a staff-only policy misses much of the risk: suppliers use their own accounts and plans, under terms the business has never seen, and receive briefs and customer lists to do the work. The business's data rule follows that data to whoever holds it.

Fix: attach the policy to each supplier agreement, state what data each may put into AI tools, and collect the terms answers for their tools or require tools from the approved list.

Verify: ask each supplier in writing which tools have held the business's data and on which plans, and add their material uses to the register with a business-side owner.

Prevent: put the AI rules into supplier onboarding, and the register rows and data held in their tools onto the offboarding list.

AI features in existing platforms under terms that changed when the feature arrived

The disclosure guide notes that AI features inside existing platforms, such as a subject-line generator or plain-language querying, are covered by the platform's existing agreement, which may have changed when the feature arrived; the evaluation guide says they need evaluating even when nothing new is bought. The data flow and integrations guide helps trace which records each feature can reach.

Fix: list the AI features switched on in each platform, read each feature's data-processing terms, and switch off any feature whose terms do not meet the rule until they do.

Verify: confirm in each admin console who can switch features on, and that the features running match the register.

Prevent: take features found switched on into the register at the playbook's weekly intake, and ask account owners to flag a new feature before anyone enables it.

Training on demos rather than the team's own tasks

The guide to building AI skills in a marketing team treats data handling as a choice everyone with a chat window open now makes every day, with a plain warning sign: exports pasted into whatever tool is open. Demonstrations use tidy examples, so nobody practises preparing real files.

Fix: run practice sessions on the team's own tasks and material that include preparing the file and choosing the approved tool, tied to the policy's data rule.

Verify: use the guide's baseline, a real task with the approved tool and its source material, and look at what each person sent as well as the output.

Prevent: repeat the baseline for new joiners and at intervals, and keep prepared examples in the shared prompt library so the safe route is the quickest to copy.

Prevention: a data rule on one line, an approved tool for each task and a quarterly check of the register

Prevention makes the safe choice the easy one and checks on a rhythm that it still is. Compress the rule to a line people can repeat: public material may go in any approved tool, personal data only in approved tools with no-training terms and only what the task needs, special-category details never. Put that line beside the tool links, in the onboarding pack and in every supplier agreement.

Give every recurring task an approved route: a tool on a plan whose terms were collected in writing, a decision that the task uses only aggregated non-personal material, or a decision to do it without AI, which is the minimisation tutorial's fallback. Purpose matters as much as content. The disclosure guide warns that records gathered to fulfil orders were never gathered so a model could draft marketing, and the privacy and consent guide traces how consent and lawful basis travel with data from system to system. Whether a new use fits that purpose is the data protection owner's question, asked before the use starts.

Keep the register honest on the governance playbook's rhythm. Each week, new requests, features found switched on and agency declarations become register rows before they go live. Each quarter, walk the whole register with use owners and agency leads, match supplier rules and platform settings to the policy, and ask each owner whether anything has gone into a tool that the row does not describe. Bring the case list to the yearly policy review, so the rule changes where the cases showed it was unclear.

Measure what the routine finds, such as the share of live uses passing the disclosure guide's checklist and the share with a current owner, not the policy's length. A quarterly walk that finds nothing deserves a second look, because the list depends on people reporting what they did. The AI troubleshooting section lists the neighbouring symptoms.

A group of dental clinics finds why patient material reached a free assistant (illustrative)

Illustrative example — not a real result

Yeskovar Dental

Yeskovar Dental, an invented group of 6 dental clinics with a marketing team of 3 and a freelance copywriter, finds patient material in a free assistant. The head of marketing tells the data protection lead that morning and starts the case list.

The list reaches 11 cases in a week. Seven are on the free assistant, 5 on personal accounts: about 140 patient reviews pasted with names to draft replies, 2 appointment exports of roughly 2,300 rows each for reactivation emails, and complaint emails, some describing treatments. Three are the freelancer's, on their own subscription with the same exports. One is a reactivation-email feature in the email platform, switched on 4 months earlier under terms nobody had read. There is no written rule and no approved tool.

The data protection lead decides what follows for the listed cases; marketing closes the gaps. The group writes the three-class rule into its policy, naming treatment details as never going into a prompt, and approves 1 drafting assistant on a business plan with no-training terms in writing. Saved export views keep first name, clinic and months since last visit; review replies are drafted with names masked and treatment details removed. The freelancer must use the approved tool and prepared files, and the email feature stays off until the lead has read its terms. At the first quarterly walk the register holds 9 owned rows; the only finding is a phone number in a notes field, caught at the sample read.

Common mistakes

  • Deciding for yourselves what the cases mean legally. Tell the data protection owner first; any duty is theirs to decide.
  • Hunting for the person instead of the gap. People hide what they expect to be blamed for.
  • Telling the tool to ignore personal details. It still receives what was pasted.
  • Trusting a training toggle in one person's settings. No-training terms belong in the contract.
  • Blocking AI with no approved alternative. The tasks still need doing, and an unchecked block is followed in name only.
  • Writing the rule for staff only. Suppliers handle the same customer data in their own tools.

Frequently asked questions

Is a case like this a data breach, and do we have to report it?

That is not a question for the marketing team to answer, and this page does not answer it. Tell whoever handles data protection for the business at once, with the list of cases as it stands: what material, which tool and plan, who and for which task. They decide what applies and what, if anything, follows, including whether anything should be deleted from the tool, recorded or reported. The marketing team's part is to hand over a complete and honest list quickly and then close the gaps that let the habit start.

Should we block AI tools until the rules are in place?

Stop the specific uses that put personal data into unapproved tools, and record that you did. A blanket block that nobody can check is followed in name only, and the tasks behind the cases still need doing. The quicker route is an approved tool on a business plan for each task the case list shows, a data rule short enough to remember and a preparation routine that beats pasting the whole export, so that the safe way is also the easy one.

Is it enough to tell the tool to ignore personal details, or to turn off training in its settings?

Neither is enough. An instruction in the prompt does not stop the tool receiving what was pasted, because whatever reaches the tool has already been processed. A training opt-out held in one person's settings can be missed by the next user and disappears with the account. The protections that hold are contractual terms for the plan the business pays for, enforced by an administrator for every account, and material reduced to the fields the task needs before anything is sent.

Terms used on this page

Frameworks

Where this page is used

Sources

  1. Regulation (EU) 2016/679 (GDPR) — EUR-Lex
  2. Direct marketing guidance — Information Commissioner's Office (UK)
  3. AI Risk Management Framework — NIST

Related